certified

This commit is contained in:
2026-06-24 21:05:06 +02:00
parent 207fcae699
commit 3c6ee9e12d
10 changed files with 610 additions and 281 deletions

View File

@@ -1,127 +1,214 @@
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.utils.data import DataLoader, RandomSampler
from torch.autograd import grad
from unlearning.Strategy import Strategy
class CertifiedRemoval(Strategy):
"""
Implements Certified Removal (Guo et al.) adapted for deep architectures
like ResNet50 by isolating and updating the final classification layer.
Implements Certified Unlearning for non-convex DNNs (Zhang et al.).
Uses a modified, stabilized stochastic Newton step using Taylor-expansion
HVP estimation across the entire parameter space, capped with calibrated noise.
"""
def __init__(self, removal_bound: float, epsilon: float, l2_reg: float = 0.1):
super().__init__()
self.removal_bound = removal_bound # gamma in the paper
self.epsilon = epsilon # Privacy budget
self.l2_reg = l2_reg # Lambda regularization term
def __init__(self, target_class_index: int, l2_reg: float = 0.0005,
gamma: float = 0.01, scale: float = 1000.0,
s1: int = 10, s2: int = 1000, std: float = 0.001, unlearn_bs: int = 2):
super().__init__(target_class_index)
self.l2_reg = l2_reg
self.gamma = gamma
self.scale = scale
self.s1 = s1
self.s2 = s2
self.std = std
self.unlearn_bs = unlearn_bs
def _get_features(self, backbone: nn.Module, loader: DataLoader, device: torch.device):
"""Passes data through the frozen ResNet backbone to extract embedding features."""
backbone.eval()
all_features = []
all_labels = []
'''
def _compute_loss_gradient(self, model, loader, device: torch.device):
model.eval()
criterion = nn.CrossEntropyLoss(reduction='sum')
params = [p for p in model.parameters() if p.requires_grad]
grad_accumulator = [torch.zeros_like(p).cpu() for p in params]
total_samples = 0
with torch.no_grad():
for inputs, labels in loader:
inputs = inputs.to(device)
# Pass through backbone to get the 2048-dimensional feature vector
features = backbone(inputs)
all_features.append(features.cpu())
all_labels.append(labels.cpu())
for data, targets in loader:
total_samples += targets.shape[0]
data, targets = data.to(device), targets.to(device)
outputs = model(data)
mini_grads = list(grad(criterion(outputs, targets), params))
for i in range(len(grad_accumulator)):
grad_accumulator[i] += mini_grads[i].cpu().detach()
for i in range(len(grad_accumulator)):
grad_accumulator[i] /= total_samples
l2_reg_term = 0.0
for param in model.parameters():
l2_reg_term += torch.norm(param, p=2)
reg_grads = list(grad(self.l2_reg * l2_reg_term, params))
for i in range(len(grad_accumulator)):
grad_accumulator[i] += reg_grads[i].cpu().detach()
return [p.to(device) for p in grad_accumulator]'''
def _compute_loss_gradient(self, model, loader, device: torch.device):
model.eval()
# Use reduction='sum' matching the original framework
criterion = nn.CrossEntropyLoss(reduction='sum')
params = [p for p in model.parameters() if p.requires_grad]
grad_accumulator = [torch.zeros_like(p).cpu() for p in params]
total_samples = 0
for data, targets in loader:
total_samples += targets.shape[0]
data, targets = data.to(device), targets.to(device)
outputs = model(data)
loss = criterion(outputs, targets)
# Incorporate L2 weight regularization directly inside the backprop graph
# to keep scaling bounded and aligned with the data volume
l2_reg_term = 0.0
for param in model.parameters():
if param.requires_grad:
l2_reg_term += torch.norm(param, p=2)
total_loss = loss + (self.l2_reg * l2_reg_term)
mini_grads = list(grad(total_loss, params, retain_graph=False))
for i in range(len(grad_accumulator)):
grad_accumulator[i] += mini_grads[i].cpu().detach()
for i in range(len(grad_accumulator)):
grad_accumulator[i] /= total_samples
return [p.to(device) for p in grad_accumulator]
def grad_batch(batch_loader, lam, model, device):
model.eval()
criterion = nn.CrossEntropyLoss(reduction='sum')
params = [p for p in model.parameters() if p.requires_grad]
grad_batch = [torch.zeros_like(p).cpu() for p in params]
num = 0
for batch_idx, (data, targets) in enumerate(batch_loader):
num += targets.shape[0]
data, targets = data.to(device), targets.to(device)
outputs = model(data)
grad_mini = list(grad(criterion(outputs, targets), params))
for i in range(len(grad_batch)):
grad_batch[i] += grad_mini[i].cpu().detach()
for i in range(len(grad_batch)):
grad_batch[i] /= num
l2_reg = 0
for param in model.parameters():
l2_reg += torch.norm(param, p=2)
grad_reg = list(grad(lam * l2_reg, params))
for i in range(len(grad_batch)):
grad_batch[i] += grad_reg[i].cpu().detach()
return [p.to(device) for p in grad_batch]
def _hvp(self, loss, params, v):
first_grads = grad(loss, params, retain_graph=True, create_graph=True)
elemwise_products = 0
for grad_elem, v_elem in zip(first_grads, v):
elemwise_products += torch.sum(grad_elem * v_elem)
# FIX 1: Set create_graph to False to prevent massive nested graph accumulation
return grad(elemwise_products, params, create_graph=False)
def _stochastic_newton_update(self, g, retain_dataset, model, device):
model.eval()
criterion = nn.CrossEntropyLoss()
params = [p for p in model.parameters() if p.requires_grad]
h_res = [torch.zeros_like(p) for p in g]
for _ in range(self.s1):
h_estimate = [p.clone() for p in g]
sampler = RandomSampler(retain_dataset, replacement=True, num_samples=self.unlearn_bs * self.s2)
res_loader = DataLoader(retain_dataset, batch_size=self.unlearn_bs, sampler=sampler)
res_iter = iter(res_loader)
for j in range(self.s2):
try:
data, target = next(res_iter)
except StopIteration:
res_iter = iter(res_loader)
data, target = next(res_iter)
data, target = data.to(device), target.to(device)
outputs = model(data)
return torch.cat(all_features, dim=0), torch.cat(all_labels, dim=0)
loss = criterion(outputs, target)
l2_reg_term = 0.0
for param in model.parameters():
l2_reg_term += torch.norm(param, p=2)
loss += (self.l2_reg + self.gamma) * l2_reg_term
h_s = self._hvp(loss, params, h_estimate)
with torch.no_grad():
for k in range(len(params)):
# FIX 2: Added .detach() to decouple history strings across iterative update blocks
#h_estimate[k] = (h_estimate[k] + g[k] - h_s[k] / self.scale).detach()
next_estimate = h_estimate[k].data + g[k].data - (h_s[k].data / self.scale)
h_estimate[k] = next_estimate.clone()
del h_s, loss, outputs
for k in range(len(params)):
h_res[k] = h_res[k] + h_estimate[k] / self.scale
return [p / self.s1 for p in h_res]
def _run(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader) -> nn.Module:
"""
Entry point expected by your Model.unlearn() architecture interface.
Applies Certified Removal strictly to the final linear layer (model.fc).
"""
'''def _run(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader) -> nn.Module:
device = next(model.parameters()).device
# Isolate the final NN (Fully connected) layer from the model
linear_head = model.fc
# Temporarily turn the fc layer into a identity pass-through
model.fc = nn.Identity()
print(">> Extracting deep features from model backbone...")
retain_features, retain_labels = self._get_features(model, retain_loader, device)
forget_features, forget_labels = self._get_features(model, forget_loader, device)
# Restore the linear head back
model.fc = linear_head
# Extract weights from the classification layer
# w shape: [num_classes, 2048]
w = model.fc.weight.data.clone().cpu()
# Compute the Exact Hessian Matrix over the remaining (retained) features
# Formula: H = (X^T * X) / N + lambda * I
N_retain = retain_features.size(0)
hessian = self._compute_hessian(retain_features=retain_features, retain_features_size = N_retain)
grad_forget = self._compute_loss_gradient(
forget_labels=forget_labels,
forget_features=forget_features,
model_weights=w)
#torch.matmul(error.t(), forget_features) / forget_features.size(0)
# Compute the Newton step update via solving: H * Delta_W^T = Grad_forget^T
delta_w = self._compute_newton_step(
tensor = hessian,
gradient= grad_forget
)
# Apply the Certified Removal update rule: W_new = W + Delta_W
new_w = w + delta_w
# Calibrate noise based on your epsilon budget
# (Guo et al. use a perturbation based on the regularization lambda and epsilon)
sigma = 2.0 / (self.l2_reg * self.epsilon)
noise = torch.randn_like(new_w) * (sigma / N_retain)
new_w = new_w + noise
# Theoretical Guarantee verification
norm_delta = torch.norm(delta_w).item()
if norm_delta > self.removal_bound:
print(f"!! Warning: Removal budget exceeded! Norm: {norm_delta:.4f} > Bound: {self.removal_bound}")
else:
print(f">> Certificate valid. Norm: {norm_delta:.4f} <= Bound: {self.removal_bound}")
# Push updated parameters back into the model instance in-place
model.fc.weight.data = new_w.to(device)
num_forget = len(forget_loader.dataset)
num_retain = len(retain_loader.dataset)
scaling_ratio = num_forget / num_retain
print(">> Certified Removal process completed successfully.")
return model
# computing the hessian matrix
def _compute_hessian(self, retain_features, retain_features_size):
print(">> Computing exact Hessian matrix...")
# N_retain = retain_features.size(0)
X_T_X = torch.matmul(retain_features.t(), retain_features)
reg_matrix = self.l2_reg * torch.eye(retain_features.size(1))
return (X_T_X / retain_features_size) + reg_matrix
def _compute_loss_gradient(self, forget_features, forget_labels, model_weights):
print(">> Calculating forget set gradients...")
num_classes = model_weights.size(0)
# Pass features through linear layer weights to get logits
logits_forget = torch.matmul(forget_features, model_weights.t())
# Apply softmax to get true class probabilities
preds_softmax = torch.softmax(logits_forget, dim=1)
print(">> Calculating base gradients over target FORGET set...")
# FIX 3: Base gradients MUST be evaluated from forget_loader to drop target class distributions
g = self._compute_loss_gradient(model, forget_loader, device)
forget_labels_one_hot = torch.nn.functional.one_hot(forget_labels, num_classes=num_classes).float()
print(">> Estimating non-convex inverse Hessian trajectories via Taylor series...")
retain_dataset = retain_loader.dataset
delta = self._stochastic_newton_update(g, retain_dataset, model, device)
print(">> Applying stabilized parameter adjustments and randomized certification noise...")
with torch.no_grad():
for i, param in enumerate(model.parameters()):
if param.requires_grad:
noise = self.std * torch.randn(param.data.size(), device=device)
#param.data.add_(-delta[i] + noise)
param.data.add_(scaling_ratio * delta[i] + noise)
print(">> Certified Unlearning process completed successfully across the complete landscape.")
return model'''
def _run(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader) -> nn.Module:
device = next(model.parameters()).device
error = preds_softmax - forget_labels_one_hot
# grad_forget shape: [num_classes, 2048]
return torch.matmul(error.t(), forget_features) / forget_features.size(0)
def _compute_newton_step(self,tensor, gradient):
print(">> Solving Newton step via system optimization...")
try:
delta_w_t = torch.linalg.solve(tensor, gradient.t())
delta_w = delta_w_t.t()
except RuntimeError:
print(">> Warning: Hessian matrix is singular. Falling back to pseudo-inverse.")
delta_w = torch.matmul(gradient, torch.linalg.pinv(tensor).t())
return delta_w
print(">> Calculating stable base gradients over the RETAIN set...")
# To match the author's snippet perfectly, g MUST be computed on the retain data.
# If this loader is too large for your VRAM, use a smaller batch size (e.g. 16 or 32)
# in your main training script when creating retain_loader.
g = self._compute_loss_gradient(model, retain_loader, device)
print(">> Estimating non-convex inverse Hessian trajectories via Taylor series...")
retain_dataset = retain_loader.dataset
delta = self._stochastic_newton_update(g, retain_dataset, model, device)
print(">> Applying parameter removal adjustments (-delta)...")
with torch.no_grad():
for i, param in enumerate(model.parameters()):
if param.requires_grad:
noise = self.std * torch.randn(param.data.size(), device=device)
# MATCHING THE SNIPPET: Subtract delta exactly as the authors do
# This removes the influence trace of the omitted data.
param.data.add_(-delta[i] + noise)
print(">> Certified Unlearning process completed successfully.")
return model